An independent open-source initiative published this weekend has found that the most widely used frontier AI models overwhelmingly default to left-leaning positions on political, social, and cultural issues.
The Neutrality Project evaluated 18 models from twelve labs across six dimensions—including civil liberties, foreign policy, speech and technology, environment, nationalism, and religion.
Of 108 measured positions, 97 landed left of center, producing an average lean of -0.41.
The strongest progressive tilt appeared on environmental questions.
Only models from xAI’s Grok family sat near the political center. None leaned to the Conservative side.
Google’s Gemma even ships a guardrail layer that actively suppresses right-leaning answers.
Unlike opaque corporate audits, the project publishes its full methodology, code, and raw results on GitHub for public scrutiny.
It uses self-anchored scales, measuring each model against its own responses rather than imposing external “neutral” definitions, making the directional findings particularly robust.
Grok 4.5, released by xAI just days earlier on July 8, emerged as the standout. Independent tests tied to the project gave it a near-perfect -0.02 bias score, the closest to zero among all models evaluated, with neutrality falling comfortably within its error band.
This performance matches the model’s explicit design focus on maximum truth-seeking over heavy-handed safety alignment.
By contrast, flagship models from OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), and Meta (Llama) consistently clustered on the progressive side.
Samuel Cardillo said on Linkedin “At a day and age where we delegate further and further our critical thinking to models trained on data and rules that we do not know about, accountability and visibility is important.”
These results echo and strengthen earlier 2025–2026 research.
A Washington Post analysis found ChatGPT delivering exclusively left-leaning arguments in roughly 80 per cent of political test questions.
Stanford researchers documented that users across parties perceive OpenAI models as having the strongest left slant.
Yale experiments showed default AI summaries subtly shifting opinions in a liberal direction.
Millions of people now consult AI for news summaries, policy explanations, voting guidance, and even personal opinions, making these left-leaning default outputs problematic.
AI models risk functioning as quiet opinion-shapers, especially among less politically engaged users.
Training data drawn heavily from the internet reflects the cultural left’s dominance in academia, media, and Big Tech content moderation.
Reinforcement learning from human feedback (RLHF) teams, often drawn from similar ideological pools, then amplifies those tendencies under the banner of “safety” and “harm reduction.”
Google has faced criticism after its artificial intelligence tool Gemini reportedly refused to show white people, preferring images of people of colour instead. https://t.co/MWp5Tlj1Pe
— Brussels Signal (@brusselssignal) February 22, 2024
